GLM-5.2 / GLM-5.3

Overview

GLM-5.2 is Z.ai's (Zhipu AI's) flagship long-horizon model in the GLM-5 family, introduced as "our latest flagship model for long-horizon tasks," marking "a substantial leap in long-horizon task capability over its predecessor GLM-5.1" and, for the first time in the family, delivering it on "a solid 1M-token context" (GLM-5.2 model card). The weight repository was created on Hugging Face on 2026-06-16 (Hugging Face model API for zai-org/GLM-5.2), matching Z.ai's own launch announcement of "GLM-5.2: Frontier Intelligence, Open Weights" with "MIT-licensed open weights" and "same API pricing as GLM-5.1" (Z.ai announcement). Its headline features are the solid 1M context, "advanced coding with flexible effort" (multiple thinking-effort levels), the IndexShare architecture change, an improved MTP layer for speculative decoding, and "Pure Open: an MIT open-source license — no regional limits, technical access without borders" (GLM-5.2 model card; GLM-5.2 README). Z.ai positions it as the highest-ranked open-source model across three long-horizon coding benchmarks (FrontierSWE, PostTrainBench, SWE-Marathon) and the strongest open-source model on standard coding benchmarks (Z.ai GLM-5.2 blog).

GLM-5.3, announced on August 14, 2026 (Yicai Global), is a post-training-only release: "It uses the same base model as GLM-5.2 — every gain comes from post-training" (Z.ai GLM-5.3 blog; Z.ai docs: GLM-5.3). Z.ai frames it as "Frontier Coding with Emergent Cyber Capabilities": a ~50% coding gain over GLM-5.2 on the in-house Z.ai Code Bench, claimed open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam (CLI), and unexpectedly fast growth in vulnerability discovery and exploitation ability (Z.ai GLM-5.3 blog; Z.ai docs: GLM-5.3). Working with security teams in China, Z.ai reports the model surfaced 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues, with an average latent lifetime of 26.6 years and the oldest flaw introduced in 1981 (Z.ai GLM-5.3 blog). GLM-5.3 is text-only, always reasoning, with a 1M-token context, 128K max output, and three effort levels (low, high, max, default max); disabling thinking is no longer supported (Z.ai docs: GLM-5.3).

Licensing/availability differs between the two. GLM-5.2 ships as MIT open weights on Hugging Face and ModelScope (Hugging Face model API; Z.ai GLM-5.2 blog). GLM-5.3 weights are not yet public at the time of writing: Z.ai states "The model weights of GLM-5.3 will be publicly available soon in two weeks," with access initially via the Z.ai API, the GLM Coding Plan, and ZCode (Z.ai GLM-5.3 blog; Z.ai docs: GLM-5.3). Notable techniques across the pair: IndexShare for cheap long-context sparse attention, SAO (with compaction) for RL on long-horizon tasks, and slime, Z.ai's open-source asynchronous RL post-training framework with Megatron for training and SGLang for rollout — GLM-5.3 reports >2.3× end-to-end RL throughput gains from system-level optimizations and training-rollout logprob agreement controlled to the 1e-7 level (Z.ai GLM-5.3 blog). Independent evaluation from Artificial Analysis placed GLM-5.3 at 60 on its Intelligence Index on August 18, 2026, level with Kimi K3 (Unite.AI).

Architecture

GLM-5.3 shares GLM-5.2's base model exactly (Z.ai GLM-5.3 blog), so the GLM-5.2 checkpoint config is the authoritative architecture source for both.

• Total parameters (checkpoint tensor count): 753,329,940,480 (~753B) (Hugging Face model API for zai-org/GLM-5.2); the family's vendor-stated headline figure is 744B total / 40B active, stated for GLM-5 on the same architecture lineage (GLM-5 model card; Z.ai docs: GLM-5)
• Active parameters per token: n.a. for GLM-5.2/5.3 specifically — not stated on any fetched vendor page (Z.ai GLM-5.2 blog; GLM-5.2 README)
• Model class: GlmMoeDsaForCausalLM, model_type: glm_moe_dsa (GLM-5.2 config.json)
• Layers: 78 hidden layers; mlp_layer_types shows the first 3 layers dense and the remaining 75 sparse (MoE) (GLM-5.2 config.json)
• Hidden size / dense FFN: hidden_size 6144, intermediate_size 12288 (GLM-5.2 config.json)
• MoE experts: 256 routed experts + 1 shared expert, 8 activated per token, expert FFN width 2048, sigmoid scoring with noaux_tc routing, routed scaling factor 2.5 (GLM-5.2 config.json)
• Attention: DSA-based sparse attention with MLA-style low-rank projections — 64 attention heads, q_lora_rank: 2048, kv_lora_rank: 512, qk_head_dim: 256 (192 nope + 64 rope), v_head_dim: 256, head_dim: 192 (GLM-5.2 config.json); DSA's lightning indexer selects top-k tokens per query, reducing core attention from O(L²) to O(Lk) while the indexer itself stays O(L²) (IndexCache paper, arXiv:2603.12201)
• IndexShare (novel component): "reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length" (GLM-5.2 model card); the indexer sits at the first of each group of 4 layers and its top-k indices are reused for all 4, removing indexer dot-product and top-k work in 3/4 of layers; trained with IndexShare from mid-training at 128K sequence length (Z.ai GLM-5.2 blog)
• IndexShare in config: index_topk_freq: 4, index_skip_topk_offset: 3, index_share_for_mtp_iteration: true, and an indexer_types list alternating one full layer followed by three shared layers across all 78 layers (GLM-5.2 config.json)
• DSA indexer sizing: 32 indexer heads, index_head_dim: 128, index_topk: 2048, interleaved indexer RoPE (GLM-5.2 config.json)
• Multi-token prediction (MTP): num_nextn_predict_layers: 1; the MTP layer was improved for speculative decoding, raising acceptance length by up to 20% — ablation: baseline 4.56 → +IndexShare+KV Share 5.10 → +rejection sampling 5.29 → +end-to-end TV loss 5.47; MTP step parameters are shared as in GLM-5.1, and in multi-step MTP the indexer runs on the first step with its top-k indices reused for later steps (GLM-5.2 config.json; Z.ai GLM-5.2 blog; GLM-5.2 model card)
• Context window: max_position_embeddings: 1048576 (1M) in the checkpoint (GLM-5.2 config.json); Z.ai API lists 1M context / 128K max output for GLM-5.2 and for GLM-5.3 (Z.ai docs: GLM-5.2; Z.ai docs: GLM-5.3). GLM-5.1's context was 200K, so 5.2 is a 5× extension (Z.ai GLM-5.2 blog)
• RoPE: default RoPE with theta raised to 8,000,000 (vs 1,000,000 in GLM-5/5.1), interleaved (GLM-5.2 config.json; GLM-5.1 config.json)
• Precision: BF16 checkpoint (dtype: bfloat16); FP8 KV-cache is used in Z.ai's large-scale agentic RL training infrastructure (GLM-5.2 config.json; Z.ai GLM-5.2 blog)
• Vocabulary: 154,880 tokens, untied embeddings (GLM-5.2 config.json)
• Effort control: GLM-5.2 exposes High and Max thinking effort (Z.ai GLM-5.2 blog); GLM-5.3 exposes low/high/max with reasoning always enabled (Z.ai docs: GLM-5.3)
• License: MIT for GLM-5.2 (Hugging Face model API; GLM-5.2 README); GLM-5.3 weights pending release (Z.ai GLM-5.3 blog)
• Serving stacks: SGLang v0.5.13.post1+, vLLM v0.23.0+, Transformers v0.5.12+, KTransformers v0.5.12+, Unsloth v0.1.47-beta+, plus vLLM-Ascend/xLLM/SGLang on Ascend NPU (GLM-5.2 model card); inference optimization builds on LayerSplit, with a throughput advantage that grows with context length (Z.ai GLM-5.2 blog)

Official vendor eval scores

Vendor-reported only. GLM-5.3 figures come from the Z.ai GLM-5.3 blog table (Z.ai ran the comparisons itself); GLM-5.2 figures from the GLM-5.2 Hugging Face card / Z.ai blog.

Nemotron 3 Ultra figures are the vendor-reported numbers from the NVIDIA Nemotron 3 Ultra model card, shown as a common reference point across model pages. n.a. means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. % better is (score − Nemotron) / Nemotron, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows — instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness. Vendor harness notes: GLM-5.2 reasoning evals use temperature 1.0 / top_p 0.95, 163,840-token generation limit, GPT-5.5 (medium) as judge, text-only subsets by default; SWE-Bench Pro uses OpenHands under a 400K context; DeepSWE uses the official pier framework with mini-swe-agent; Terminal-Bench 2.1 Claude Code scores are averaged over 5 runs; FrontierSWE was evaluated by Proximal, PostTrainBench by PostTrainBench, and SWE-Marathon by Abundant AI, all at 1M context and max effort (GLM-5.2 model card). GLM-5.3's HLE-with-tools run used a 300K context with context management and GPT-5.6-luna (medium) as judge (Z.ai GLM-5.3 blog).

Benchmark explanations

• Terminal-Bench 2.1 / 3.0 — real-world command-line tasks solved autonomously in a sandboxed terminal; 3.0 is the harder successor generation, described by Z.ai's docs as measuring complex task completion in a real terminal environment (Z.ai docs: GLM-5.3; GLM-5.2 model card).
• DeepSWE v1.1 — long-horizon software-engineering task suite run with the official pier framework and mini-swe-agent harness, each task in an isolated no-internet container (GLM-5.2 model card).
• SWE-bench Pro — contamination-resistant GitHub issue resolution requiring repo-level patches, run with OpenHands under a 400K context (GLM-5.2 model card).
• SWE-Marathon — very long-horizon software engineering evaluated by Abundant AI at 1M context, testing sustained multi-hour work (GLM-5.2 model card).
• FrontierSWE — frontier-difficulty software-engineering tasks scored as a "dominance" rate, evaluated by Proximal at 1M context and max effort (GLM-5.2 model card).
• NL2Repo — generate an entire working repository from a natural-language specification, with rule- and LLM-based anti-cheat checks (GLM-5.2 model card).
• ProgramBench — 200 hard programming instances run through Claude Code with up to 2,000 turns and 6-hour timeouts; "Almost Solved" is the stricter partial-credit tier (GLM-5.2 model card; Z.ai GLM-5.3 blog).
• PostTrainBench — machine-learning post-training research tasks (run the experiment, improve a model) evaluated by the benchmark authors at 1M context (GLM-5.2 model card).
• Agents' Last Exam (ALE-CLI) — real-world professional agent scenarios emphasizing cross-tool collaboration and long-horizon execution, scored pass@1 in a CLI harness (Yicai Global; Z.ai GLM-5.3 blog).
• Toolathlon Verified — verified multi-tool "chores" requiring correct orchestration of many external tools (Z.ai GLM-5.3 blog).
• AutomationBench v1.0.6 — end-to-end automation of real workflow tasks by an agent (Z.ai GLM-5.3 blog).
• CyberGym — starts from white-box source code and tests whether the model can identify and validate vulnerabilities by triggering faults (Z.ai GLM-5.3 blog).
• ExploitBench — requires deeper reasoning about real vulnerabilities and producing working exploitation, i.e. further up the exploitation chain than CyberGym (Z.ai GLM-5.3 blog).
• ExploitGym — counts how many exploitation tasks a model completes under time-normalized budgets (2-hour and 6-hour), normalized by per-model throughput (Z.ai GLM-5.3 blog).
• Humanity's Last Exam (HLE) — broad expert-written exam of frontier knowledge and reasoning; the "with tools" variant permits search/tool use (GLM-5.2 model card; Z.ai GLM-5.3 blog).
• GDPval-AA v2 — Elo-style evaluation of real-world, economically valuable knowledge work spanning 44 occupations (Yicai Global; Z.ai GLM-5.3 blog).
• AIME 2026 / HMMT / IMOAnswerBench — competition mathematics (American Invitational Mathematics Examination, Harvard-MIT Mathematics Tournament, IMO-style problems) graded on exact final answers (GLM-5.2 model card).
• GPQA-Diamond — hardest subset of graduate-level, Google-proof science multiple-choice questions (GLM-5.2 model card).
• CritPt — frontier physics-reasoning research problems, reported alongside HLE in the reasoning block (GLM-5.2 model card).
• MCP-Atlas — 500-task public subset of Model Context Protocol tool-calling tasks, LLM-judged (Gemini 3.0 Pro as judge) (GLM-5.2 model card).
• Tool-Decathlon — ten-domain agentic tool-use battery run through the official evaluation service with 128K max tokens (GLM-5.2 model card).
• Z.ai Code Bench — Z.ai's private benchmark placing coding agents in complex local development environments, scoring end-to-end task completion and checklist accuracy plus output-token efficiency (Z.ai GLM-5.3 blog).